JP6661839B1 - Time series data diagnosis device, additional learning method, and program - Google Patents

Abstract

The time-series data diagnostic device (10) includes a storage unit (11), a dissimilarity calculation unit (13), a diagnosis result information generation unit (14), an evaluation information acquisition unit (16), and an update unit (17). And A dissimilarity calculation unit (13) calculates a dissimilarity between the time-series data to be diagnosed and the learned sample subsequence data stored in the storage means. The diagnosis result information generation unit (14) generates diagnosis result information of the time-series data of the diagnosis target based on the difference degree calculated by the difference degree calculation unit (13). An evaluation information acquisition unit (16) acquires evaluation information indicating an evaluation as to whether the diagnosis result information is a proper diagnosis result. The update unit (17) stores the evaluation information acquisition unit (16) based on the time-series data of the diagnosis target corresponding to the diagnosis result information when the evaluation information acquisition unit (16) acquires the evaluation information indicating the evaluation that the diagnosis result is not appropriate. The learned sample subsequence data stored in (11) is updated.

Description

  The present invention relates to a time-series data diagnosis device, an additional learning method, and a program.

  Various time-series data obtained over time are accumulated in equipment such as buildings and factories, equipment on factory lines, automobiles, railway vehicles, and the like. A device that diagnoses the time-series data and detects an abnormality of the device has been studied. In addition, studies have been made to improve the diagnostic accuracy by learning time-series data of these devices in normal or abnormal times. For example, Patent Literature 1 discloses a technique of generating sample subsequence data from statistical values of similar subsequence data among subsequence data extracted from time-series data as a learned model to be compared in diagnosis. It has been disclosed. By using the generated sample subsequence data as a comparison target for diagnosis, the accuracy and the amount of calculation are balanced.

International Publication No. 2016/117086 JP-A-2006-033778

  It is desirable that sample subsequence data to be a trained model be generated with high accuracy from the beginning. However, due to time or economic constraints, the accuracy is not always perfect from the beginning. It is also necessary to cope with changes in various situations during operation, such as aging of equipment and seasonal fluctuations. Therefore, it is desirable that the sample subsequence data can be updated by additionally learning even during operation. On the other hand, Patent Literature 1 does not disclose a method of additional learning for updating a learned model during operation. On the other hand, Patent Literature 2 discloses a technique in which time series data diagnosed by the diagnosis unit as having no sign of failure is added to a cluster in a time zone to which the time series data belongs, and the cluster is updated. . However, the technique described in Patent Literature 2 is a technique for updating a cluster based on the added time-series data and a plurality of time-series data which are members before the addition, and thus requires a large amount of calculation. Therefore, it cannot be said that it is suitable for additional learning during operation that requires processing speed. Also, if the accuracy of learning is significantly reduced by simply setting a small amount of calculation, it becomes impossible to withstand actual operation.

  The present invention has been made in view of the above circumstances, and has as its object to provide a time-series data diagnosis device, an additional learning method, and a program that can perform additional learning with a small amount of calculation and high accuracy.

In order to achieve the above object, a time-series data diagnosis apparatus of the present invention includes a storage unit, a difference calculation unit, a diagnosis result information generation unit, an evaluation information acquisition unit, and an update unit. The storage means includes intermediate sample subsequence data generated from statistical values of subsequence data of each group obtained by grouping time series data to be learned with similar temporally continuous subsequence data, and intermediate sample subsequence data. And the learned sample subsequence data generated based on the above , and parameters including the first threshold value . The dissimilarity calculating means calculates the dissimilarity between the time-series data to be diagnosed and the learned sample subsequence data stored in the storage means. The diagnosis result information generation unit compares the difference calculated by the difference calculation unit with the first threshold to generate diagnosis result information of the time-series data to be diagnosed. The evaluation information obtaining means obtains evaluation information indicating whether or not the diagnosis result information is a valid diagnosis result. The updating means, when the degree of difference is equal to or greater than the first threshold, and when the evaluation information obtaining means obtains evaluation information indicating an evaluation that is not a valid diagnostic result, the diagnostic result information evaluated as not being a valid diagnostic result Generating an intermediate sample subsequence data for update from the statistical value of the subsequence data of each group grouped by similar subsequence data that is temporally continuous from the time series data of the diagnosis target corresponding to Sample subsequence data is generated based on the column data and the intermediate sample subsequence data stored in the storage unit, and the learned sample subsequence data stored in the storage unit is updated with the generated sample subsequence data.

  According to the present invention, there is provided a time-series data diagnostic apparatus that can perform additional learning with a small amount of calculation and high accuracy by updating learned sample subsequence data generated based on time-series data to be learned. Can be provided.

The figure showing an example of the learned model construction method according to Embodiment 1 of the present invention. Functional block diagram of a time-series data diagnostic device according to Embodiment 1 of the present invention FIG. 2 is a hardware configuration diagram of the time-series data diagnostic device according to the first embodiment of the present invention. Flowchart of time-series data diagnosis processing according to Embodiment 1 of the present invention The figure showing an example of the updating method of the trained sample subsequence data according to Embodiment 1 of the present invention. Functional block diagram of a time-series data diagnostic device according to Embodiment 2 of the present invention Flowchart of time-series data diagnosis processing according to Embodiment 2 of the present invention

(Embodiment 1)
Hereinafter, an embodiment in which a time-series data diagnosis device, an additional learning method, and a program of the present invention are applied to online learning during operation will be described with reference to the drawings.

  The time series data to be diagnosed is based on the passage of time from various sensors such as buildings, factories, etc., factory line equipment, thermometers, vibration meters, voltmeters, etc. attached to automobiles, railway vehicles, etc. These are various data obtained. Then, the time-series data diagnosis device acquires time-series data, diagnoses the data in real time, and outputs a diagnosis result in order to monitor the states of these devices.

  In practice, a learned model is constructed by a learning device before using the time-series data diagnosis device. Then, the time-series data diagnosis device diagnoses the time-series data transmitted in real time with reference to the learned model. Further, the time-series data diagnosis device updates the learned model by performing additional learning during operation in order to increase the accuracy of diagnosis.

  First, a method of constructing a learned model as a premise of the present embodiment will be described with reference to FIG.

  An example is shown in which a learning device different from the time-series data diagnosis device according to the present embodiment constructs a learned model. However, the time-series data diagnosis device may double as the learning device. Further, in the present embodiment, the learning device acquires normal time-series data and generates a learned model, so that the time-series data diagnostic device determines that the time-series data similar to the generated learned model is normal. Diagnose.

  The time-series data to be learned is digital data in which time and data are associated with each other and continuous values are converted into discrete data at a specific sampling rate. The learning device first extracts W temporally continuous data from the acquired time-series data of the learning target. The W data thus extracted is referred to as subsequence data, and W is referred to as the subsequence data length. Then, the learning device sequentially generates the plurality of partial sequence data while slightly shifting the time at which the time-series data is to be extracted.

  Next, the learning device groups the time-series data into similar partial sequence data that are temporally continuous. Here, since the time-series data obtained from the device is usually a value having continuity, the fact that the temporally continuous sub-sequence data tend to be similar to each other makes use of the temporally continuous data. The sub-sequence data to be grouped. Further, the subsequence data including W data can be regarded as a point in the W-dimensional space. Therefore, the difference between the partial sequence data is represented by the Euclidean distance between two points in the W-dimensional space, and the longer the distance is, the larger the difference is, indicating that the partial sequence data are not similar to each other.

  As illustrated in FIG. 1, the learning device groups the subsequence data whose distance from the reference subsequence data is equal to or smaller than a specific threshold. Then, when partial string data whose distance from the leading partial string data exceeds a specific threshold value appears, the learning apparatus repeats the same processing with the partial string data as a new leading partial string data. This threshold is set to ε / 2 by using a sampling error upper limit ε which is an upper limit of an approximation error allowed for a sample generated from a statistical value. The sampling error upper limit ε is set in advance using the statistical value of the time series data. Since the distance between the subsequence data belonging to the group and the head subsequence data is all ε / 2 or less, the radius ε / 2 on the W dimension centered on the point representing the head subsequence data is obtained. Since all the subsequence data belonging to the same group are included inside the sphere, the distances between arbitrary subsequence data belonging to the same group are all less than or equal to ε.

  Next, the learning device generates intermediate sample subsequence data by averaging the grouped subsequence data. The intermediate sample subsequence data is represented by a statistical value of W data included in all subsequence data belonging to the group. This statistical value may be an average value such as an arithmetic mean or a geometric mean, or may be another statistical value. Typically, in the W-dimensional space, the point representing the intermediate sample subsequence data may be the center of gravity of each subsequence data. For example, if the arithmetic mean is adopted as the statistical value, the partial sequence data A (a, b, c), the partial sequence data B (d, e, f), and the partial sequence data C (g, h, i) The intermediate sample subsequence data of the group including the three subsequence data is ((a + d + g) / 3, (b + e + h) / 3, (c + f + i) / 3).

  Next, the learning device calculates the feature amount of each partial sequence data, and sorts the partial sequence data in order of the feature amount. The feature value of the subsequence data may be an average value of W data included. This sorting is a process utilizing the fact that time-series data acquired from a device usually has periodicity. In this way, the learning device can arrange partial sequence data that are not temporally continuous but are similar to each other close to each other.

  Next, the learning device groups similar intermediate sample subsequence data from the sorted plurality of intermediate sample subsequence data by a process similar to the process of generating the intermediate sample subsequence data described above. Then, the learning device generates learned sample subsequence data from the statistical values of the grouped intermediate sample subsequence data.

  Further, the learning device newly acquires time-series data as trial data. Then, the learning device calculates the degree of difference between the trained sample subsequence data and the trial data, and calculates an average value m of the degree of difference, a standard deviation σ, and the like as statistical values. Then, the learning device determines a threshold value Tha and a threshold value Thb of the degree of difference as parameters for diagnosis based on these statistical values. The substring data length W is also included in the parameter for diagnosis.

  Here, the threshold Thb, which is the second threshold, is larger than the threshold Th, which is the first threshold. For example, it is assumed that Tha = 50 and Thb = 70 are set. In this case, in the diagnosis result information generation processing described later, time series data having a degree of difference from the learned sample partial sequence data of Tha, that is, 50 or more, is determined to be abnormal. Further, in the additional learning processing described later, when reflecting the time series data having a difference of Thb from the learned sample subsequence data of 70 or more to the learned sample subsequence data, the additional generation generated from the time series data is performed. The sample subsequence data for training is added to the trained sample subsequence data. On the other hand, when reflecting the time series data having a degree of difference from the learned sample subsequence data of Thb, that is, less than 70, to the learned sample subsequence data, the learned sample subsequence data is updated.

  In this way, the learning device generates learned sample subsequence data and parameters for diagnosis as a learned model.

  Next, a functional configuration of the time-series data diagnostic device 10 according to the present embodiment will be described with reference to FIG.

  The time-series data diagnostic device 10 includes a storage unit 11 for storing various data, a time-series data acquisition unit 12 for acquiring time-series data, a dissimilarity calculation unit 13 for generating dissimilarity information, and generating diagnostic result information. A diagnostic result information generating unit 14 that outputs diagnostic result information, a diagnostic result information output unit 15 that outputs diagnostic result information, an evaluation information acquiring unit 16 that acquires evaluation information, an updating unit 17 that updates learned sample subsequence data, An addition unit 18 for adding already-sampled subsequence data and a parameter update unit 19 for updating parameters are provided.

  The storage unit 11 stores a learned model generated by the above-described learning device, that is, learned sample subsequence data, and parameters including a threshold value Tha and a threshold value Thb of the degree of difference, a subsequence data length W, and the like. . The storage unit 11 also stores intermediate sample subsequence data, which is intermediate data generated by the learning device. Further, the storage unit 11 also stores trial data used when the learning device generates the parameters.

  The time-series data acquisition unit 12 acquires the time-series data to be diagnosed from the time-series data transmission device 30.

  The dissimilarity calculating unit 13 calculates a dissimilarity between the acquired time-series data and the learned sample subsequence data stored in the storage unit 11. Specifically, the dissimilarity calculating unit 13 extracts a plurality of partial sequence data having a partial sequence data length W from the time-series data, and calculates a distance between the extracted plurality of partial sequence data and the trained sample partial sequence data. . Then, the closest distance among the calculated distances is set as the degree of difference.

  The diagnosis result information generation unit 14 compares the difference generated by the difference calculation unit 13 with a threshold Tha included in the parameters stored in the storage unit 11 to generate diagnosis result information. Specifically, when the degree of difference from the learned sample subsequence data is equal to or greater than the threshold Tha, the diagnosis result information generation unit 14 generates diagnosis result information indicating that the time-series data to be diagnosed is abnormal, When the degree of difference from the trained sample subsequence data is less than the threshold value Tha, diagnostic result information indicating that the time-series data to be diagnosed is normal is generated.

  The diagnosis result information output unit 15 outputs the diagnosis result information generated by the diagnosis result information generation unit 14. Specifically, the diagnosis result information output unit 15 displays the diagnosis result information on a display described later. Alternatively, the diagnostic result information output unit 15 may transmit the diagnostic result information to an external device, system, or the like.

  The evaluation information obtaining unit 16 obtains evaluation information indicating whether or not the diagnosis result information is a valid diagnosis result. Specifically, the person in charge who browsed the diagnosis result information output by the diagnosis result information output unit 15 comprehensively judges the time-series data of the diagnosis target, the actual state of the device, the past experience, and the like. Evaluate whether the result information is valid. Then, the evaluation information acquisition unit 16 acquires the evaluation information input by the person in charge operating the keyboard described later. The evaluation that the diagnosis result information is not valid includes the case where the diagnosis result is normal but the diagnosis result is abnormal, and the case where the diagnosis result is abnormal but the diagnosis result is normal is valid. , Is included. The evaluation information acquisition unit 16 may acquire the evaluation information from an external device, system, or the like.

  When the difference calculated by the difference calculating unit 13 is equal to or greater than the threshold Tha, and the evaluation information obtaining unit 16 obtains evaluation information indicating an evaluation that is not a valid diagnosis result, the updating unit 17 adds the evaluation information to the diagnosis result information. From the corresponding time-series data, a plurality of partial sequence data having temporally different degrees of difference of ε / 2 or less are grouped. Then, the updating unit 17 generates update intermediate sample subsequence data from the statistical value of the subsequence data belonging to each group. Then, the updating unit 17 combines the intermediate sample subsequence data for update and the intermediate sample subsequence data stored in the storage unit 11 and further groups those having a difference degree of ε / 2 or less. Then, the updating unit 17 generates sample subsequence data from the statistical values of the intermediate sample subsequence data for update belonging to each group. Further, the updating unit 17 updates the learned sample subsequence data stored in the storage unit 11 with the generated sample subsequence data.

  When the difference calculated by the difference calculating unit 13 is equal to or greater than the threshold Tha and the evaluation information obtaining unit 16 obtains evaluation information indicating an evaluation that is not a valid diagnosis result, the adding unit 18 determines whether the evaluation result is not a valid diagnosis result. The additional sample subsequence data is generated based on the corresponding time series data. Then, the adding unit 18 adds additional sample subsequence data to the learned sample subsequence data stored in the storage unit 11.

  The parameter updating unit 19 updates the parameters based on the updated sample subsequence data when the evaluation information obtaining unit 16 obtains evaluation information indicating an evaluation that is not a valid diagnosis result.

  Next, a hardware configuration of the time-series data diagnostic device 10 will be described with reference to FIG. The time-series data diagnostic device 10 includes a CPU (Central Processing Unit) 101 that executes various processes, a RAM (Random Access Memory) 102 that is a volatile memory, and a ROM (Read Only Memory) 103 that is a nonvolatile memory. A hard disk drive 104 for storing various information, a keyboard 105 for receiving operations, a display 106 for displaying information, and a network card 107 for transmitting and receiving information.

  The CPU 101 reads various programs stored in the hard disk drive 104 into the RAM 102 and executes them to execute various processes described below.

  The RAM 102 is a volatile memory and is used as a work area of the CPU 101.

  The ROM 103 is a non-volatile memory and stores a control program for the basic operation of the time-series data diagnostic device 10 executed by the CPU 101, a basic input output system (BIOS), and the like.

  The CPU 101, the RAM 102, and the ROM 103 cooperate to function as the above-described dissimilarity calculation unit 13, diagnosis result information generation unit 14, update unit 17, addition unit 18, and parameter update unit 19.

  The hard disk drive 104 is a medium for storing information, and stores the various data described above. The hard disk drive 104 functions as the storage unit 11.

  The keyboard 105 is an information input device, and receives input of various information by a user operation. The keyboard 105 functions as the evaluation information acquisition unit 16.

  The display 106 is an information display device such as a liquid crystal panel and a projector. The display 106 displays information necessary for the user to operate. The display 106 functions as the diagnosis result information output unit 15.

  The network card 107 is a communication device that communicates information with an external device, and is communicably connected to the time-series data transmission device 30. The network card 107 functions as the time-series data acquisition unit 12. In some cases, the network card 107 also functions as the diagnosis result information output unit 15 or the evaluation information acquisition unit 16.

  Next, the operation of the time-series data diagnostic device 10 will be described with reference to FIG.

  The storage unit 11 of the time-series data diagnostic device 10 stores learned sample subsequence data as a trained model and parameters including a threshold value Tha and a threshold value Thb of the degree of difference, a subsequence data length W, and the like. I have. The storage unit 11 also stores intermediate sample subsequence data and trial data. Therefore, the storage unit 11 functions as a storage unit that stores the intermediate sample subsequence data, the learned sample subsequence data, and the parameters.

  When the user of the time-series data diagnosis device 10 inputs a request to start the time-series data diagnosis process by operating the keyboard 105, the time-series data acquisition unit 12 of the time-series data diagnosis device 10 The time series data of the diagnosis target is acquired from 30 (step S11).

  Next, the dissimilarity calculating unit 13 calculates the dissimilarity between the time-series data acquired in step S11 and the learned sample subsequence data stored in the storage unit 11 (step S12). Specifically, the dissimilarity calculating unit 13 extracts a plurality of partial sequence data having a partial sequence data length W from the time-series data. Then, the distance between the extracted plurality of partial sequence data and the learned sample partial sequence data is calculated. Then, the dissimilarity calculating unit 13 determines the closest distance among the calculated distances as the dissimilarity between the time-series data and the trained sample subsequence data. In step S12, the dissimilarity calculating unit 13 functions as a dissimilarity calculating unit that calculates the dissimilarity between the time-series data to be diagnosed and the learned sample subsequence data stored in the storage unit 11. Step S12 functions as a dissimilarity calculation step for calculating the dissimilarity.

  Subsequently, the diagnosis result information generation unit 14 compares the degree of difference calculated in step S12 with the threshold Tha to generate diagnosis result information (step S13). Specifically, when the degree of difference from the learned sample subsequence data is equal to or greater than the threshold Tha, the diagnosis result information generation unit 14 generates diagnosis result information indicating that the time-series data of the diagnosis target is abnormal, When the degree of difference from the trained sample subsequence data is less than the threshold value Tha, diagnostic result information indicating that the time-series data to be diagnosed is normal is generated. In step S13, the diagnosis result information generation unit 14 compares the difference calculated by the difference calculation unit 13 with the first threshold value, and generates diagnosis result information of time-series data to be diagnosed. Functions as generating means. Step S13 functions as a diagnostic result information generating step of generating diagnostic result information.

  Next, the diagnosis result information output unit 15 outputs the diagnosis result information generated by the diagnosis result information generation unit 14 (Step S14). Specifically, the diagnosis result information output unit 15 displays the diagnosis result information on the display 106. Alternatively, the diagnostic result information output unit 15 may transmit the diagnostic result information to an external device, system, or the like via the network card 107.

  Next, the evaluation information acquisition unit 16 acquires evaluation information indicating whether or not the diagnosis result information is a valid diagnosis result (step S15). Specifically, the person who views the diagnosis result information output in step S14 comprehensively determines the time series data of the diagnosis target, the actual device status, past experience, and the like, and the diagnosis result information is displayed. Evaluate whether the results are appropriate. Then, the evaluation information acquisition unit 16 acquires the evaluation information input by the person in charge operating the keyboard 105. The evaluation information acquisition unit 16 may acquire the evaluation information from an external device, system, or the like. In step S15, the evaluation information obtaining unit 16 functions as evaluation information obtaining means for obtaining evaluation information indicating whether or not the diagnosis result information is a valid diagnosis result. Step S15 functions as an evaluation information obtaining step of obtaining evaluation information.

  Subsequently, the time-series data diagnosis device 10 determines whether or not the evaluation information acquired in step S15 indicates an evaluation that is not a valid diagnosis result (step S16). More specifically, the time series data is evaluated as abnormal with respect to the diagnosis result that the time series data is normal, and the time series data is evaluated as abnormal with respect to the diagnosis result that the time series data is abnormal. If the evaluation is that the data is normal, it is determined that the evaluation information indicates an evaluation that is not a valid diagnosis result.

  If the time-series data diagnosis device 10 determines that the evaluation information acquired in step S15 indicates an evaluation that is not a valid diagnosis result (step S16: Yes), the difference calculated in step S12 is stored in the storage unit 11. It is determined whether it is less than the threshold Thb included in the stored parameters (step S17). On the other hand, if the time-series data diagnosis device 10 determines that the evaluation information acquired in step S15 indicates an evaluation indicating a valid diagnosis result (step S16: No), the time-series data diagnosis process ends.

  When the time-series data diagnostic device 10 determines that the difference is less than the threshold Thb (step S17: Yes), the updating unit 17 updates the learned sample subsequence data (step S18). Specifically, the updating unit 17 specifies the learned sample subsequence data having a degree of difference less than Thb as a target of the update process. Then, as shown in FIG. 5, the updating unit 17 generates the intermediate sample subsequence data by the same processing as the method of constructing the learned model. The generated intermediate sample subsequence data is referred to as update intermediate sample subsequence data. Then, the update unit 17 combines the intermediate sample subsequence data stored in the storage unit 11 corresponding to the learned sample subsequence data to be updated, and the generated intermediate sample subsequence data for update together with these. Sort in order of feature value.

  Then, the updating unit 17 groups similar intermediate sample subsequence data from the plurality of sorted intermediate sample subsequence data. The updating unit 17 generates updated learned sample subsequence data from the statistical values of the grouped intermediate sample subsequence data. In step S18, the updating unit 17 functions as an updating unit that updates the learned sample partial sequence data stored in the storage unit 11 with the generated sample partial sequence data. Step S18 functions as an update step of updating the learned sample subsequence data.

  Returning to FIG. 4, when the time-series data diagnostic device 10 determines that the difference is equal to or greater than the threshold Thb (step S17: No), the adding unit 18 adds the partial sequence data to the learned sample partial sequence data ( Step S19). Specifically, the adding unit 18 generates additional sample subsequence data from the time-series data corresponding to the diagnosis result information in the same procedure as when constructing the learned model. Then, the adding unit 18 adds the generated additional sample subsequence data to the learned sample subsequence data stored in the storage unit 11. In step S19, the adding unit 18 functions as an adding unit that adds additional sample subsequence data to the learned sample subsequence data stored in the storage unit 11.

  Subsequent to step S18 or step S19, the parameter updating unit 19 updates parameters for diagnosis stored in the storage unit 11 (step S20). Specifically, the parameter updating unit 19 determines the difference between the updated or added learned sample subsequence data and the trial data stored in the storage unit 11 in the same manner as when constructing the learned model. Is calculated, and the average value m, the standard deviation σ, and the like of the degree of difference are recalculated as the statistical values. Then, the parameter updating unit 19 updates the difference thresholds Tha and Thb, which are parameters for diagnosis, based on these statistical values. In step S20, the parameter updating unit 19 functions as a parameter updating unit that updates a parameter based on the updated learned sample subsequence data.

  According to the time-series data diagnostic device 10 according to the present embodiment, the trained model is not reconstructed from the time-series data to be subjected to the additional learning and the time-series data used for building the trained model. The trained sample subsequence data can be generated by utilizing the intermediate sample subsequence data generated when the model is constructed. Therefore, the amount of calculation can be reduced, and it is suitable for additional learning, especially for online learning to learn during operation. In addition, since the time series data used for additional learning is not temporally continuous with the time series data used for building the trained model, the accuracy is hardly changed compared to rebuilding the trained model. The result is obtained. Therefore, additional learning can be performed with a small amount of calculation and high accuracy.

  In addition, the time-series data diagnostic device 10 according to the present embodiment distributes the update of the learned sample subsequence data by the update unit 17 and the addition to the learned sample subsequence data by the addition unit 18 based on the threshold Thb. . By this, if there is a trained sample subsequence data similar to the time series data to be additionally trained, this is updated, and if there is no similar trained sample subsequence data, the trained sample subsequence data is newly updated. Will be added. Therefore, it is possible to perform appropriate additional learning without excessively increasing the sample subsequence data that has already been learned.

(Embodiment 2)
In the first embodiment, an example has been described in which the learned sample subsequence data is updated or added when the evaluation information indicating that the diagnosis result information indicating the abnormality is normal is acquired. In the present embodiment, an example in which learned sample subsequence data is deleted when evaluation information indicating abnormality is acquired for diagnosis result information indicating normality is obtained. Hereinafter, a description will be given focusing on portions different from the first embodiment.

  As shown in FIG. 6, the time-series data diagnostic device 10 according to the present embodiment further includes, in addition to the functional units described in the first embodiment, a deletion unit 20 that deletes learned sample subsequence data.

  The deletion unit 20 responds to the diagnosis result when the difference calculated by the difference calculation unit 13 is less than the threshold Tha and the evaluation information obtaining unit 16 obtains evaluation information indicating that the evaluation is not a valid diagnosis result. The learned sample subsequence data whose degree of difference from the time series data to be performed is less than the threshold Tha is deleted from the storage unit 11.

  More specifically, the operation of the time-series data diagnostic device 10 will be described with reference to the drawings. As illustrated in FIG. 7, the time-series data diagnosis device 10 determines that the evaluation information acquired in step S15 indicates an evaluation that is not a valid diagnosis result in step S16 of the time-series data diagnosis processing (step S16: Yes), it is determined whether or not the degree of difference calculated in step S12 is equal to or greater than a threshold Tha (step S21). When the time-series data diagnostic device 10 determines that the degree of difference is equal to or greater than the threshold Tha (step S21: Yes), the process proceeds to step S17, which is a process of determining whether the degree of difference is less than the threshold Thb. The same processing as in the first mode is performed.

  On the other hand, if the time-series data diagnostic device 10 determines that the degree of difference is less than the threshold Tha (step S21: No), the deletion unit 20 determines the intermediate value corresponding to the learned sample subsequence data whose degree of difference is less than the threshold Tha. The sample subsequence data is deleted (step S22). In step S12, since the dissimilarity is calculated using the combination having the smallest difference between the time-series data and the learned sample subsequence data, the learned sample subsequence data whose dissimilarity from the time-series data is less than the threshold Tha Must be present at least one or more times. Therefore, specifically, the deletion unit 20 calculates the degree of difference from the intermediate sample subsequence data corresponding to the learned sample subsequence data whose degree of difference is less than the threshold Tha. Then, the deletion unit 20 deletes the intermediate sample subsequence data whose calculated degree of difference is less than the threshold Tha. Furthermore, the deletion unit 20 updates the learned sample subsequence data based on the intermediate sample subsequence data that has not been deleted. In step S22, the deletion unit 20 deletes the intermediate sample subsequence data whose degree of difference from the time-series data to be diagnosed is less than the threshold Tha from the storage unit 11, and based on the remaining intermediate sample subsequence data. It functions as a deletion unit that updates the learned sample subsequence data stored in the storage unit 11.

  Subsequent to step S22, the parameter update unit 19 updates the parameters based on the updated learned sample subsequence data as in the first embodiment (step S20).

  The time-series data diagnostic device 10 according to the present embodiment updates the learned sample subsequence data when acquiring the evaluation information indicating abnormality with respect to the diagnosis result information indicating normal. This further improves the accuracy of the learned sample subsequence data.

  Further, the time-series data diagnostic device 10 according to the present embodiment updates the learned sample subsequence data by deleting the intermediate sample subsequence data. Therefore, the amount of calculation can be reduced as compared with deleting the time series data itself and deleting the intermediate sample subsequence data. Further, additional learning can be performed with higher accuracy than when the learned sample subsequence data is deleted.

(Modification)
The present invention is not limited to the above-described embodiment, and various other modifications are possible.

  In the above-described embodiment, an example is shown in which a trained model is constructed from time-series data of a normal device, and a normal diagnosis is performed when the degree of difference from the trained sample subsequence data included in the trained model is small. Was. However, the scope of the present invention is not limited to this, and a trained model is constructed from the time-series data of the abnormal device, and when the degree of difference from the trained sample subsequence data included in the trained model is small, it is diagnosed as abnormal. The configuration may be as follows. In this case, the present invention can be applied as it is by reading the terms “normal” and “abnormal” in the above-described embodiments in the opposite directions.

  In the above-described embodiment, the time-series data diagnostic device 10 calculates the degree of difference based on the Euclidean distance in the W-dimensional space. However, this is an example, and the degree of difference may be calculated by another calculation method. Further, the time-series data diagnosis device 10 may also serve as a learning device.

  Although the example in which each of the dissimilarity, the threshold value Tha, and the threshold value Thb is one numerical value has been described, these may be a combination of a plurality of numerical values. In that case, the threshold should be interpreted as a condition for a plurality of numerical values.

  The time-series data diagnostic device 10 according to the present embodiment can be realized using a normal computer without using a dedicated device. For example, the time-series data diagnosis device 10 that executes the above-described processing may be configured by installing the program for executing any of the above-described programs into the computer from a recording medium that stores the program. Further, one time-series data diagnosis device 10 may be configured by a plurality of computers operating in cooperation.

  The method for supplying the program to the computer is arbitrary. For example, the information may be supplied via a communication line, a communication network, a communication system, or the like.

  When a part of the above functions is provided by an OS (Operating System), a part other than the functions provided by the OS may be provided by a program.

  The present invention is capable of various embodiments and modifications without departing from the broad spirit and scope of the present invention. Further, the above-described embodiment is for describing the present invention, and does not limit the scope of the present invention. That is, the scope of the present invention is shown not by the embodiments but by the claims. Various modifications made within the scope of the claims and the equivalents of the invention are considered to be within the scope of the present invention.

  10 time series data diagnosis device, 11 storage unit, 12 time series data acquisition unit, 13 difference degree calculation unit, 14 diagnosis result information generation unit, 15 diagnosis result information output unit, 16 evaluation information acquisition unit, 17 update unit, 18 addition Unit, 19 parameter updating unit, 20 deleting unit, 30 time series data transmission device, 101 CPU, 102 RAM, 103 ROM, 104 hard disk drive, 105 keyboard, 106 display, 107 network card.

Claims (7)

Based on the intermediate sample subsequence data generated from the statistical value of the subsequence data of each group obtained by grouping the time series data to be trained with similar temporally continuous subsequence data, and the intermediate sample subsequence data Storage means for storing the learned sample subsequence data generated by the above, and a parameter including a first threshold value ;
Time-series data of a diagnosis target, and a difference calculation unit that calculates a difference between the learned sample subsequence data stored in the storage unit,
A diagnosis result information generation unit that compares the difference calculated by the difference calculation unit with the first threshold, and generates diagnosis result information of the time-series data of the diagnosis target;
Evaluation information acquisition means for acquiring evaluation information indicating whether the diagnosis result information is a valid diagnosis result,
When the degree of difference is equal to or greater than the first threshold, and the evaluation information obtaining unit obtains evaluation information indicating evaluation that is not a valid diagnosis result, the diagnosis result information evaluated as not being a valid diagnosis result Generating an intermediate sample subsequence data for update from the statistical value of the subsequence data of each group grouped by similar subsequence data successive in time from the time- series data of the diagnosis object corresponding to the Generating sample subsequence data based on the intermediate sample subsequence data and the intermediate sample subsequence data stored in the storage means, and storing the learned sampled sequence stored in the storage means with the generated sample subsequence data Updating means for updating the sample subsequence data,
Time series data diagnostic device. When the degree of difference is equal to or greater than the first threshold and the evaluation information obtaining unit obtains evaluation information indicating an evaluation that is not a valid diagnosis result, the time series of the diagnosis target corresponding to the diagnosis result information Generating additional sample subsequence data based on the data, further comprising an additional unit that adds the additional sample subsequence data to the learned sample subsequence data stored in the storage unit;
The time-series data diagnostic device according to claim 1 . The parameter further includes a second threshold greater than the first threshold;
The updating means updates the learned sample subsequence data when the dissimilarity is less than the second threshold value,
The adding unit adds the generated additional sample subsequence data when the degree of difference is equal to or greater than the second threshold.
The time-series data diagnostic device according to claim 2 . When the degree of difference is less than the first threshold value and the evaluation information acquisition unit acquires evaluation information indicating an evaluation that the evaluation result is not a valid diagnosis result, the time-series data of the diagnosis target corresponding to the diagnosis result Of the intermediate sample subsequence data corresponding to the learned sample subsequence data whose difference degree was less than the first threshold value, the difference degree with the time-series data of the diagnosis target corresponding to the diagnosis result is The intermediate sample subsequence data that is less than the first threshold is deleted from the storage unit, and the learned sample subsequence data stored in the storage unit is updated based on the remaining intermediate sample subsequence data. Further comprising:
The time-series data diagnostic device according to any one of claims 1 to 3 . A parameter updating unit that updates the parameter based on the updated learned sample subsequence data when the evaluation information obtaining unit obtains evaluation information indicating an evaluation that the diagnosis result information is not a valid diagnosis result. Further comprising,
The time-series data diagnostic device according to any one of claims 1 to 4 . Executed by the time-series data diagnostic device,
Based on the time-series data of the diagnosis target and the intermediate sample sub-sequence data generated from the statistical value of the sub-sequence data of each group obtained by grouping the time-series data of the learning target with similar temporally continuous sub-sequence data. When the evaluation information indicating the evaluation that the degree of difference from the learned sample subsequence data generated by the above is not appropriate is acquired, the data is grouped by similar subsequence data that is temporally continuous from the time series data of the diagnosis target . Generating an intermediate sample subsequence data for update from the statistical value of the subsequence data of each group,
Wherein the intermediate samples subsequence data for updating, and the intermediate samples subsequence data to generate sample subsequence data based on,
Wherein the generated specimen subsequence data, updating the learned sample subsequence data,
Additional learning method. On the computer,
Based on the time-series data of the diagnosis target and the intermediate sample sub-sequence data generated from the statistical value of the sub-sequence data of each group obtained by grouping the time-series data of the learning target with similar temporally continuous sub-sequence data. A difference calculation step of calculating a difference between the learned sample subsequence data generated and
A diagnosis result information generation step of comparing the difference calculated in the difference calculation step with a first threshold to generate diagnosis result information of the time-series data of the diagnosis target;
An evaluation information obtaining step of obtaining evaluation information indicating whether the diagnosis result information is a valid diagnosis result or not;
When the degree of difference is equal to or greater than the first threshold and the evaluation information obtaining step obtains evaluation information indicating an evaluation that is not a valid diagnosis result, the diagnosis result information evaluated as not a proper diagnosis result Generating an intermediate sample subsequence data for update from the statistical value of the subsequence data of each group grouped by similar subsequence data successive in time from the time- series data of the diagnosis object corresponding to the An updating step of generating sample subsequence data based on the intermediate sample subsequence data and the intermediate sample subsequence data, and updating the learned sample subsequence data with the generated sample subsequence data;
A program for executing

Images